DataPacts

Seller track · Lesson 6 of 9 · 9:07

The four terms that decide what you give away

What you'll learn

  • How exclusivity shuts other buyers out
  • Why a perpetual grant never ends
  • What happens to a trained model
  • Who pays when a claim comes

Free lesson

Enter your email to watch this lesson and the rest of the course.

Transcript

Here's a number to start with. Of forty-five company data programs, only twenty-four publish enough contract terms to compare. The other twenty-one publish nothing, or leave every clause to the agreement you sign later. So the contract, not the sales page, is where you find out what you're giving away. Let's go through the terms that matter most.

Most of the risk sits in five clauses: the grant, deletion, trained models, the indemnity and the liability cap. We'll fold them into four terms. Exclusivity. How long the grant lasts. What happens to your data and the models. And who pays when a claim comes.

Why the contract and not the program page? A clause in a published master agreement binds more than the same words on a FAQ or a landing page. Of the programs in one dated comparison, only a handful publish full contract text. The rest come from FAQs and marketing pages.

Term one, exclusivity.

An exclusive grant shuts out every other buyer for the scope it covers. A non-exclusive grant leaves room for others, though a warranty can still limit parallel sales. So the label alone matters less than what the clause lets you do next, and whether your data can be licensed again.

Published terms show three models. One data labeling company says its licenses are exclusive by default. One data program grants nonexclusively, with a no-prior-grant warranty in section 5.1. And one talent platform says it alone sells a record for ninety days.

Exclusivity also works as a screen. One talent platform's listing says data already licensed exclusively through a broker is not a fit. And one data marketplace states that ninety-five percent of its licensors signed exclusively.

So what does exclusivity pay? A genetics company disclosed fifty million dollars from a drug maker to extend exclusivity in drug-target discovery. For company data, no public premium figure exists. So don't assume an exclusive grant comes with one.

Watch for pages that disagree. One program's marketing pages use the word exclusively, while its master terms say nonexclusive. And exclusivity and audit can sit only in the order form, if the order form adds them. That means the master terms are a starting point, never the whole deal.

Term two, how long it lasts.

The term decides how long the buyer may use the data. Published terms range from a few years to perpetual and irrevocable. A big social platform's filing describes access arrangements of two to three years.

Two more limits sit beside the term. The permitted purpose lists what the buyer may do with the data, and anything outside it is a breach. One filing grants data for the sole purpose of developing and training a foundation model. And most published grants are worldwide, so any restriction usually names who may receive the data, not where.

Four patterns appear. Perpetual and irrevocable. A fixed or agreed period, such as three years in one illustrative draft. Until one side terminates. And pages that say two different things.

The same buyer can answer twice. One data marketplace's customer agreement runs until terminated, while its own FAQ says the license agreement is perpetual. Until terminated means someone can end it. Perpetual means nobody has to.

Duration also depends on who can pass the data on. One marketplace's license is non-sublicensable. One data program's grant is sublicensable through multiple tiers, and assignable in a merger, though not to a competitor.

And ending your contract doesn't end theirs. One marketplace's licensees keep access twelve months after termination, and academic licensees three years. A code marketplace states that completed licenses survive.

Term three, deletion and the models.

A deletion clause says when the buyer destroys its copies of your raw data. In one program's master terms that's the earlier of a hundred and twenty days after delivery, or termination, with backups excluded. Notice what it covers: the raw data. It rarely reaches the models already trained.

Deadlines differ a lot between buyers. Some set a fixed date, thirty days after delivery or after the project ends. A kill clause forces deletion within thirty days, or deletion comes on request. Others delete at the end of the term, or if no deal is made. And some have no duty to delete at all.

Of twenty-four programs, none promises to delete or retrain a model already trained on your data. Every published deletion clause covers the raw data only. One data labeling company puts it plainly: models already trained on the data cannot be unlearned.

What gets built from your data usually belongs to the buyer. One program's terms give the licensee ownership of derivatives, models, weights, outputs and synthetic data. In one public filing it runs the other way: the data company receives a license to the model its partner trains.

A regulator can order what a contract won't. The FTC required a photo app developer to delete the models it built from users' photos and videos. That's a remedy in an enforcement action, not a contract term. And under GDPR, erasure duties apply whatever a license says.

Term four, indemnity and the cap.

Three clauses work as a chain. In the warranty, you promise you may license the data and that consents exist. The indemnity enforces that promise: you defend and pay the buyer's third-party claims. The cap then limits how much anyone can recover.

In published terms, it's the seller. One program's licensor covers IP, privacy, publicity and data protection claims, in section 8.1. A marketplace's sellers cover regulator claims too. And a code marketplace states an uncapped ownership warranty.

Caps are usually a multiple of fees. One program caps at three times fees, except for fraud, gross negligence and willful misconduct. Another caps at two times the fees of the last twelve months. And no published survey of typical caps exists.

In the EU, the Data Act makes imposed limits for intent or gross negligence always unfair. Unfair terms that one company imposes on another aren't binding, for contracts after September twelfth, 2025. Whether any of it reaches you is a question for a lawyer, so this isn't legal advice.

The governing law clause picks the law and the forum, and published buyer terms choose the buyer's home state. One program picks California law and arbitration. Others pick Washington, Delaware or New York. For you, that can mean a foreign forum and its costs.

Confidentiality is the quiet term. In one program's master terms it lasts three years, and indefinitely for trade secrets. These clauses protect the deal terms and the data, and they survive termination for a fixed period. So what you learn in the deal stays under wraps long after it ends, for both sides.

Last, before you sign.

One more trap sits in delivery. In one program's master terms the buyer has thirty business days to reject your data. After that, silence counts as approval and the data is deemed accepted. A further sixty days cover latent defects. So the clock is running before you've noticed.

Ask for the order form first, before you share any data. Negotiate it. Limit the grant by time or purpose. And remember that an empty row in a comparison is a question to ask, not a sign the terms are worse.

And not signing is a valid outcome. You can keep the data and license nothing at all. Declining to sign is always on the table, whatever the pressure to decide fast, and a no now can still become a yes later, once your rights are in order.

So, in one line. Check the exclusivity, ask how long the grant lasts, remember that deletion never reaches a trained model, and know who pays when a claim comes. Read the contract, not the program page, and have a lawyer in your jurisdiction review it before you sign.

Next up is lesson A7, Getting more than one offer.

Read and watch the full lesson free. Enter your email to unlock every lesson of the academy.

Sources on DataPacts